Adaptive Trellis-Coded Modulation over Predicted Flat Fading Channels Sorour Falahati , Arne Svensson , Mikael Sternad and Hong Mei Department of Signals and Systems, Uppsala University, P.O. Box 528, SE-751 20 Uppsala, Sweden. Email: Sorour.Falahati, Mikael.Sternad@signal.uu.se Department of Signals and Systems, Chalmers University of Technology, SE-412 96 G¨ oteborg, Sweden. Email:
[email protected] Department of Information Technology, Uppsala University, P.O. Box 337, SE-751 05 Uppsala, Sweden. Email:
[email protected]
Abstract— We consider the optimum design of an adaptive scheme based on TCM and predicted CSI for flat Rayleigh fading channels, intended for fast link adaptation. The question of how to optimally adjust the data rate to maximize the spectral efficiency, subject to a BER constraint when imperfect CSI is taken into account, is answered. An optimum solution based on the predicted SNR and the prediction error variance is derived. The performance of the adaptive TCM scheme is illustrated by utilizing seven 4-D trellis codes based on the International Telecommunications Union’s ITU-T V.34 modem standard. The results indicate that the gain in spectral efficiency by adaptive TCM is about 1-2 dB compared to adaptive uncoded M-QAM, for most of average SNRs. Finally, the performance of the adaptive TCM and in particular, its loss at high SNRs compared to adaptive uncoded M-QAM, are investigated in details.
I. I NTRODUCTION The problems of spectrum utilization in the mobile environment stimulate a great deal of research, especially in the area of adaptive transmission. The purpose of adaptive transmission is to adjust the signaling parameters as the channel varies, with the best possible usage of resources. Recently, adaptive transmission based on uncoded M-ary Quadrature Amplitude Modulation (M-QAM) has gained considerable attention [1]– [4]. To improve the obtainable spectral efficiency, more complex schemes would be required. A competitive candidate for this purpose is Trellis-Coded Modulation (TCM) [5]–[8]. The reason is that a TCM scheme has the potential of attaining a target Bit Error Rate (BER) at a lower average Signal-to-Noise Ratio (SNR), as compared to an uncoded M-QAM with the same spectral efficiency. The adaptive transmission techniques often rely on the Channel State Information (CSI) being available at the transmitter. Perfect CSI is usually assumed in the design of adaptive modulation schemes. However, in order for the scheme to meet the expectations in practice, realistic assumptions should be used in the design [1], [2], [4], [8], [9]. An important requirement is that the CSI often needs to be available a few slots ahead to setup the transmission. For very slowly fading Hong Mei contributed in this work during her studies at Chalmers University of Technology in G¨oteborg, Sweden. This work was funded by the Swedish Foundation for Strategic Research.
channels, the outdated estimates of CSI are still useful for the adaptation purposes. However, for fast fading channels, the feedback delays result in inaccurate CSI estimates at the time of transmission. Hence, it is logical to perform fast link adaptation based on channel predictions. The optimum adaptive TCM scheme considered here, adjusts the data rate such that the spectral efficiency is maximized subject to a target BER and constant average transmit power. The scheme takes the statistical information of predicted Rayleigh fading channels into account. This approach distinguishes the present work from other solutions found in the literature (see e.g. [5]–[7]). The channel predictions utilized in this study are obtained through an unbiased quadratic regression of the past noisy channel estimates [10]. The statistical information is exploited to improve the overall system performance and to fulfill the BER requirements even in presence of the prediction error. The spectral efficiency of the adaptive TCM is evaluated and compared with the adaptive uncoded M-QAM. The results show that the expected gain in spectral efficiency based on TCM is obtained for a large but limited range of average SNR. At high SNRs, superior performance by the uncoded M-QAM is observed. Detailed investigations show that this phenomenon is caused by the impact of the prediction quality on the rate boundaries as well as the higher rate transmission of uncoded M-QAM. In the following, Section II describes the system model, the structure of the TCM scheme and the statistical characteristics of the channel predictor. The optimal rate adaptation is derived in Section III under a BER constraint. Analytical results are presented in Section IV and are discussed in detail. Finally, Section V summarizes the results. II. S YSTEM MODEL TCM schemes with different constellation sizes are provided at the transmitter. The channel is modeled by a flat Rayleigh fading channel with AWGN. At the receiver, a channel estimator estimates the channel gain which is used to perform coherent detection of the data bits. A time-series of these channel estimates are used for prediction purposes. Based on
the predicted channel and the prediction quality, a constellation size is selected and reported to the transmitter via the feedback channel. There, the adaptive trellis encoder maps the incoming data bits to the symbols belonging to the selected constellation where constant average power is assumed. Let denote the received signal sampled at time where is the symbol period and index represents the signal is given by sample at the symbol rate. Here,
(1)
where is the transmitted symbol with average transmit power . Moreover, is a sample of a complex AWGN with zero mean and time-invariant variance . Finally, is the complex channel gain with circular Gaussian distribution with zero mean and a variance denoted by . The channel gain power denoted by , which is proportional to the instantaneous received SNR, is then distributed. The auto-correlation function of the channel gain is determined by the Doppler spectrum. The average received SNR is given . For the constant transmit power assumed by here, the instantaneous received SNR is and the predicted instantaneous received SNR is where is the predicted instantaneous channel gain power. Since the channel estimation error is believed to have a minor effect on the performance compared to the prediction error, perfect channel estimation is considered here for the demodulation. Also, to perform decoding at the rate used by the encoder at any time instant, an error free feed-back channel is assumed. Moreover, the prediction horizon is presumed to be long enough to take the computational and signaling delays in the adaptation control loop into account. The proposed adaptive TCM scheme utilizes a set of twodimensional ( -D) trellis-codes with different constellation sizes, similar to [7], where is a positive integer. Let denote the number of constellations available at the transmitter, bits per symbol for each of size with
. A -D signal constellation is obtained by -fold Cartesian products of 2-D signal constellations. At every th transmission time instant, coded bits are mapped onto symbols, each from a 2-D constellation of size for
. These symbols are sent during consecutive symbol intervals. At the receiver, soft decoding based on the Viterbi algorithm is performed where the squared Euclidean distance is used as metric in the decoder. The channel gain is predicted using an unbiased quadratic power predictor proposed in [10] which is optimal in the Mean Square Error (MSE) sense. For this type of predictor, the error variance of the predicted channel gain, denoted by , is related to the error variance of the predicted power, denoted by , as [11]
(2)
In this work, is used as a measure indicating the prediction quality, where the higher it is, the less reliable the predicted values become. In section 8 of [11], it is shown that if an
optimal unbiased power predictor is used which provides a given , then the pdf of the instantaneous true SNR , conditioned on the instantaneous predicted SNR , will be
(3)
where is the Heaviside’s step function and is the zeroth order modified Bessel function. The time index was dropped in the pdf expression since and are both stationary random processes. Also, the pdf of is given by
(4) which is a shifted -distribution, with the shift
caused by the bias compensation.
III. AVERAGE SPECTRAL EFFICIENCY OF ADAPTIVE TCM The spectral efficiency of a (coded) modulation scheme is given by the average data rate per unit bandwidth. Let denote the instantaneous data rate when constellation size is chosen, for
, and denote the transmitted signal bandwidth. Hence, the average spectral efficiency is simply given by ·½ (5) where , denote the rate region , with boundaries. These thresholds are defined as the ranges of the instantaneous predicted SNR over which different constella , tions are used by the transmitter, i.e. if coded bits per symbol are transmitted. In other words, a signal constellation or equivalently a transmission rate is assigned to each rate region boundary. The task here is to find the optimum rate region boundaries such that the average spectral efficiency is maximized subject to two constraints. The first constraint considered in this study is the BER constraint given by
(6) where is the instantaneous BER of the modulation
with constellation size of in terms of the instantaneous predicted SNR, and denotes the target BER. This constraint guarantees that the average BER does not exceed the target BER. The second constraint is the use of constant average transmit power. Power adaptation is not included, since the gain achieved in spectral efficiency by rate adaptation is higher than by using power adaptation (at least for adaptive M-QAM) [3] and it increases the complexity. To find the optimal rate region boundaries based on (6), an is needed. This expression is obtained expression for by averaging , the instantaneous BER as a function
TABLE I
PARAMETERS
AND
0
10 OF
TCM CODES . −1
30.9487
11.4545
15.4067 91.3050
-105.9332
9.8150
6.7158
-77.7363
8.4824
-194.7515
89.9720
8.2248
120.4751 303.5932 192.5175
75.5856
9.1826
5.8901
62.6452
5.6452
7.7474
104.3384
7.1502
-168.6707
8.9040
72.3208
9.6999
8.1244
-184.9618
8.5936
65.0278
9.4727
8.6210
-375.4897
8.8549
74.4199
9.8427
8.3420
-295.9806
8.7070
103.9613
9.3817
−2
10
1 2 3 4 5 6 7
10
M=256 M=128
−3
10
M=64 −4
M=32
10
−5
M=4
10
M=8
M=16
−6
10
of the instantaneous true SNR, , on an AWGN channel, over all values of , i.e.
(7) where is given by (3). Hence, what remains is to determine which will be discussed next. Tight approximations for the TCM BER performance on AWGN channels at high SNR, based on analytical or numerical methods, are available in the literature [7]. However, to perform the integration in (7), an expression which approximates the BER well enough over the whole range of SNRs is required. This expression is preferred to be in a form that based on (7). results in a closed form expression for This simplifies the search for optimal thresholds. In this study, considering the formulation of given by (3), a general expression given by [12] (8)
assures fulfilling the above mentioned requirements. Here, the parameters
and , are real numbers where the latter set takes only non-negative values. Both sets are determined by using curve fitting with least squares techniques utilizing the simulated BER for different constellation sizes on AWGN channels. Examples are shown in Section IV. Consequently, using (3) and (8) in (7), leads to
¾ ¾ ¾ ¾ ¾
¾
(9)
The optimal rate region boundaries, , are found by solving (6) with equality where (9) is used to approximate . An analytical expression for the optimal rate region boundaries can be found for
. If , numerical approaches such as the bisection method [13] can be used to find the optimal thresholds. Once they are evaluated, the spectral efficiency can easily be obtained according to (5).
0
5
10
15
[dB]
20
25
30
Fig. 1. BER performance of TCM codes on AWGN channels where M denotes the constellation size. The curves are approximations obtained from (8) and the circles are outcomes of simulation.
IV. R ESULTS AND DISCUSSION The illustrative example considered here is given for the 4-D TCM (i.e. ) with 16 states, utilizing a rate convolutional encoder and a bit mapper similar to those used in the International Telecommunication Union’s ITU-T V.34 modem standard. Nyquist pulse shaping, i.e.
, and 2-D signal constellations of sizes , corresponding to 4-QAM, 8STAR, 16-QAM, 32-CROSS, 64-QAM, 128-CROSS and 256QAM, respectively, are presumed. At each even time instant, information bits are encoded to coded bits which are mapped onto two symbols, each from a 2-D constellation of size . These symbols are transmitted during the next two time instants and result in a instantaneous data rate given by . The BER performance of the above mentioned trellis-codes on AWGN channels was simulated for various SNRs. As explained in Section III, the simulation outcomes were used to approximate the BER for AWGN channels based on (8). The resulting parameters are listed in Table I. The BER approximations are depicted in Fig. 1 and compared with the simulated BERs for different constellation sizes. It is clear that the approximations are tight within 1.5 dB in comparison with the simulated results. Based on these approximations, the performance of the adaptive TCM scheme for flat Rayleigh fading channels with average power , is evaluated. The results are compared to those of the uncoded adaptive M-QAM scheme [4] utilizing the same signal constellations as the adaptive TCM but having the instantaneous data rate given by . Fig. 2 shows the average spectral efficiency, , versus the average SNR, , for and , and prediction and . The results demonstrate error variances that in a large range of average SNR, the adaptive TCM somewhat outperforms the adaptive uncoded M-QAM (by 12 dB). The spectral efficiency curves flatten out due to the limited number of data rates.
−2
6 4
10
[bps/Hz]
8 TCM QAM
−4
10
TCM QAM
2 0 0
−6
10
20
[dB]
30
40
10
15
20
25
[dB]
30
35
6 4
[bps/Hz]
8 TCM QAM
−3
10
2 −4
0 0
10
10
20 [dB]
30
40
Fig. 2. The average spectral efficiency of the adaptive transmission based on TCM and uncoded M-QAM schemes versus average received SNR. The upper and lower plots correspond to and , respectively. The solid and dashed lines correspond to and , respectively.
An interesting phenomenon observed here is that the gain by TCM gradually vanishes as the average SNR increases. This effect is more pronounced for small s and large prediction error variances. At high average SNRs, the transmitter often transmits at the highest data rate which happens to be larger for uncoded M-QAM than TCM, and hence, provides a higher spectral efficeincy. In addition, there is a non-intuitive effect due to the influence of the prediction quality on the location of the rate region boundaries. Fig. 3 displays how
vary as increases. the SNR thresholds for At dB, the adaptive TCM has lower thresholds as compared to uncoded M-QAM and achieves a higher average spectral efficiency as shown in Fig. 2. However, at dB, the uncoded M-QAM exhibits lower thresholds for smaller constellations and higher spectral efficiency. To explain why the rate region boundaries behave in this way, the integrands and outcome of the integral in (7) are illustrated graphically in Fig. 4. On AWGN channels for a given SNR, as shown in Fig. 4(a), lower BER is obtained by TCM as compared to uncoded QAM, except for a limited range of low SNRs (this is typical for coded modulations). After performing the integration in (7), this feature is preserved for the predicted BER. Comparing Fig. 4-(a) and 4-(c) shows: TCM begins to outperform uncoded QAM at a lower BER as compared to the case for AWGN channels. It is observed (but not shown here) that the predicted SNR value corresponding to the starting point of better performance of TCM than uncoded M-QAM, increases as the tail of the pdf in (3) becomes heavier. This tail is strongly influenced by the ratio between the predicted and average SNRs. In a fading dip, i.e. when , the predictor basically fails in making a good prediction and results in a large prediction error variance relative to the predicted SNR. When the temporary channel condition improves , i.e. when , the channel prediction accordingly becomes more accurate. This yields a smaller error variance relative to the
TCM QAM 28
30
32
[dB]
34
36
Fig. 3. Instantaneous BER versus instantaneous predicted SNR of the adaptive transmission based on TCM and uncoded M-QAM schemes for and . The upper and lower plots correspond dB and dB, respectively. to
predicted value. Fig. 4-(b) shows how the shape of varies for different values of the average SNR, given and are constant. At dB, the pdf which is centered around dB and has small tails, shows the channel quality is above the average and the prediction of dB is close to the actual value. For dB, the conditional dB with heavier pdf becomes more spread around tails, meaning that the prediction of dB contributes in more errors. Finally, at dB where the predicted SNR is far below the average, the tail of the pdf below dB, grows substantially. This indicates that the predicted value is unreliable and the channel experiences a fading dip. More examples of this property are given in Fig. 5, which shows how the average SNR affects some of the rate boundaries of adaptive uncoded M-QAM and TCM accordingly, for the given and . Finally Fig. 6 illustrates how the prediction quality influences the data rate of the adaptive scheme. The spectral efficiency decreases as the predictor degrades, as demonstrated in Fig. 2. This stems from the fact that the rate boundaries are increased as the prediction error variance increases as displayed in Fig. 6, where the average SNR and remain constant. Also, a larger variation in thresholds for smaller constellations is observed here. This is simply because higher data rates are used for channels that are predicted to be in favorable conditions. In those cases, the prediction errors can hardly change the channel status and impose a lower data rate. Note that since the above conclusions are drawn based on the relative BER performance of TCM and uncoded M-QAM, they remain valid also for better approximations (compare the simulation results in Fig. 1 here with the BER performance in Fig. 2 of [3]). V. C ONCLUSIONS The optimum design of an adaptive scheme based on TCM that utilizes imperfectly predicted CSI for flat Rayleigh fading
0
40
10
TCM QAM
TCM QAM
35
−2
10
30 M=32
M=4
[dB]
M=256
25
−4
10
0
10
20
(a)
−3
4
x 10
[dB]
40
15
M=32
M=256
2
10 5
1 0 0
10
20
(b) 0
30
[dB]
40
Fig. 5.
10
15
20
[dB]
25
30
35
40
Optimum rate region boundaries in terms of the average SNR for and . The diagonal line corresponds to .
10
M=4
20
SNR=20 dB SNR=30 dB SNR=40 dB
3
30
TCM QAM −1
10 −5
10
M=4
TCM QAM
M=32 M=256
10
28
30
32
34 36 (c) [dB]
38
40
42
¾
−10
M=256
−2
10
Fig. 4. Plot (a) illustrates the instantaneous BER on AWGN channels versus instantaneous SNR. Plot (b) shows the conditional pdf given by (3), for dB and . Plot (c) displays the instantaneous BER versus instantaneous predicted SNR for dB and .
M=32 M=4 −3
10
channels, is investigated. Assuming constant average transmit power, the data rate is adjusted based on the predicted SNR and the prediction error variance to maximize the spectral efficiency and satisfy a BER requirement. The adaptive TCM is compared to an adaptive uncoded M-QAM scheme from the performance point of view. The results indicate that the former increases the spectral efficiency for a large but limited range of average SNR as compared to the latter. At high average SNR, uncoded M-QAM outperforms TCM due to the higher data rate or (and) the lower thresholds for smaller constellations. The impact is more pronounced for small s and large prediction error variances. The SNR thresholds above the average SNR are not affected much by prediction errors, while those below the average are raised considerably. This effect increases with the prediction error variance. R EFERENCES [1] A. J. Goldsmith and S. Chua, “Variable-rate variable-power MQAM for fading channels,” IEEE Transactions on Communications, vol. 45, no. 10, pp. 1218–1230, Oct. 1997. [2] M.-S. Alouini and A. J. Goldsmith, “Adaptive M-QAM modulation over Nakagami fading channels,” Kluwer Wireless Personal Comm., vol. 13, pp. 119–143, May 2000. [3] S. T. Chung and A. J. Goldsmith, “Degrees of freedom in adaptive modulation: a unified view,” IEEE Transactions on Communications, vol. 49, no. 9, pp. 1561–1571, Sept. 2001. [4] S. Falahati, A. Svensson, M. Sternad, and T. Ekman, “Adaptive modulation systems for predicted wireless channels,” in Proc. IEEE Global Telecommunications Conference, San Francisco, USA, Dec. 2003.
5
10
15
[dB]
20
25
30
Fig. 6. Optimum rate region boundaries in terms of the prediction error variance for and dB.
[5] S. M. Alamouti and S. Kallel, “Adaptive trellis-coded multiple-phaseshift keying for Rayleigh fading channels,” IEEE Transactions on Communications, vol. 42, no. 6, pp. 2305–2314, June 1994. [6] A. J. Goldsmith and S. Chua, “Adaptive coded modulation for fading channels,” IEEE Transactions on Communications, vol. 46, no. 5, pp. 595–602, May 1998. [7] K. J. Hole, H. Holm, and G. E. Øien, “Adaptive multidimensional coded modulation over flat fading channels,” IEEE Journal on Selected Areas in Communications, vol. 18, no. 7, pp. 1153–1158, July 2000. [8] V. K. N. Lau and M. D. Macleod, “Variable-rate trellis coded QAM for flat-fading channels,” IEEE Transactions on Communications, vol. 49, no. 9, pp. 1550–1560, Sept. 2001. [9] D. L. Goeckel, “Adaptive coding for time-varying channels using outdated fading estimates,” IEEE Transactions on Communications, vol. 47, no. 6, pp. 844–855, June 1999. [10] T. Ekman, M. Sternad, and A. Ahl´en, “Unbiased power prediction on broadband channel,” in Proc. IEEE Vehicular Technology Conference, Vancouver, Canada, Sept. 2002. [11] T. Ekman, “Prediction of mobile radio channels, modeling and design,” PhD thesis, Signals and Systems, Uppsala University, Uppsala, Sweden. http://www.signal.uu.se/Publications/abstracts/a023.html, Oct. 2002. [12] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and Products, Academic Press, San Diego, CA, 5th edition, 1994. [13] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical Recipes in C, Cambridge, U.K., 2nd edition, 1997.
